Evaluating Scale Transferability and Observation-Based Calibration in High-Resolution PM2.5 Downscaling

High-resolution PM2.5 mapping is increasingly required for exposure assessment and for assessing compliance with environmental policies. However, statistical downscaling from coarse gridded products assumes that PM2.5-predictor relationships are transferable across spatial resolutions. This study evaluated that assumption and developed a two-stage framework for 100 m PM2.5 mapping in Zhejiang Province, China. In Stage 1, the 1 km CHAP-PM2.5 product was downscaled with multi-source predictors using XGBoost and TabICL v2. In Stage 2, monitoring observations were integrated through daily bias correction and residual calibration. Scale transferability was assessed using coefficient-side similarity, output-side transfer tests, and a simulation-based recoverability experiment; calibration was evaluated using stratified and buffered site-based cross-validation. Scale transferability was only partial: predictor effects and transferred outputs became less stable as the resolution gap increased, and downscaling recovered only part of the known sub-kilometre heterogeneity. Stage 1 mainly enhanced spatial texture, whereas Stage 2 produced clearer improvements in agreement with ground observations and recovered much of the variability lost in the coarse-resolution data product. In cross-validation, the best calibrated product increased R2 from 0.55 to 0.74 and reduced RMSE from 7.3 to 5.5 μg/m3. This paper provides an extensive evaluation for 2020 and provides a series of quality-assured daily 100 m resolution PM2.5 maps for Zhejiang province for 2020–2024 which can be flexibly aggregated in space and time for different applications.